| Direction of Arrival(DOA)estimation is an important aspect of array signal processing,widely used in radar,sonar,and home intelligent speech assistants.With the successful application of deep learning in the field of computer vision,more and more scholars are using deep learning frameworks to solve the DOA estimation problem in array signal processing.Deep learning-based DOA estimation algorithms are data-driven and do not rely on array mathematical model assumptions.Therefore,compared to model-based algorithms,they are expected to be more robust in combating array model errors.However,in the presence of array errors or Gaussian spatial correlated noise(Gaussian spatial colored noise),existing DOA estimation algorithms based on deep learning have poor generalization ability and application effectiveness.Therefore,the following two DOA estimation algorithms are proposed for different scenarios.1.When the array phase error is unknown and may change in different environments,the array phase error result in differences between the array error in training and the array error in the testing stage.To address this issue,a robust DOA estimation algorithm based on deep neural networks,called Magnitude-based Deep Neural Network(shorten as MDNN),is proposed to utilize the amplitude characteristics of the received signals in the array.The DOA estimation accuracy of this algorithm is independent of the array phase error.In addition,to further improve the accuracy of DOA estimation while preserving the independence of array phase errors,the MDNN-WF algorithm is proposed by combining the MDNN algorithm with the traditional joint iterative(WF)algorithm.The simulation results in different scenarios indicate that compared with DOA estimation algorithms based on deep neural networks,the MDNN algorithm is more robust to array phase errors.In addition,the results indicate that the MDNNWF algorithm inherits the robustness of the MDNN algorithm and achieves higher estimation accuracy.2.Under the background of Gaussian spatial correlation noise,the performance of the traditional DOA estimation algorithm based on the second-order statistics is degraded,and the higher-order cumulant has the ability to naturally suppress the Gaussian noise.Therefore,traditional DOA estimation algorithms based on fourth-order cumulants have good performance in Gaussian spatial correlation noise backgrounds,but their performance deteriorates in low signal-to-noise ratio and limited number of samples.To address these issues,a DOA estimation method combining covariance and convolutional neural networks was studied,labeled as the COV-CNN algorithm,which uses the real and imaginary parts of covariance as inputs to the convolutional neural network.In addition,in order to improve the robustness to Gaussian spatial correlation noise,a DOA estimation algorithm is proposed,which combines fourth order cumulants and convolutional Neural Network(CNN),shorten as FOC-CNN.The proposed FOC-CNN algorithm uses,the real and imaginary parts of the fourth order cumulants as inputs to the convolutional neural network.This algorithm performs well under low SNRs and limited number of samples,and improves the performance under Gaussian spatial correlation noise.At the same time,experimental data was used to verify the corresponding algorithm. |
